Watch it remember: cross-session agent memory in ~300 lines

I built a Virtual CFO. It’s an AI agent that gives a small business financial advice and remembers that business across separate sessions. Ask it whether to buy a new oven this quarter and it recalls a cash crunch the owner mentioned months earlier in a different chat, then writes a new memory from today’s conversation.
Live, I’ll run the working app and show the agent recall a memory it was never given in this session, with the retrieved memories and their similarity scores on screen. Then I’ll go into the database, run the actual vector search, open a memory document with its embedding, and show the one index field that keeps each business’s memory isolated. I’ll switch to a second business to prove the isolation holds. I’ll also walk the agent loop: embed the question, search that business’s memory, pull its operational data, let the model reason, write a new memory back. About 300 lines of Python.